AI Agent Security Incidents: WorkOS, Binance and Controls

Overview: AI Agent Security Incidents

Explore AI agent security incidents, distinguishing the WorkOS and Binance cases and the controls discussed in CrowdStrike’s response.

The AI Agent Security Problem: What the WorkOS and Binance Breaches Show, and Why CrowdStrike Built a Kill Switch

Reviewed by Imran Khan Pathan, Editor at AI Tech Safar. This report is built from WorkOS’s own security bulletin, Binance’s published findings as reported by CNBC, Reuters and others, blockchain-analytics attributions from TRM Labs, Chainalysis and Elliptic, and CrowdStrike’s press release and event coverage. We checked each claim against more than one source and flag what comes from a company’s own account. Our first draft contained errors that we corrected: it described the Binance theft as a case of a rogue agent (it was a human attack that used a company AI tool as a stepping stone), it said both incidents happened “in a single week,” it left out that the Binance attack began in June, it called CrowdStrike’s launch a reaction to these breaches (it was announced on November 3, before WorkOS’s full update), and it cited several sources we could not verify.

Last updated: November 9, 2026.

AI agent security incidents at WorkOS and Binance, with CrowdStrike's kill-switch response.

TL;DR

  • WorkOS: A security researcher’s AI agent, built on the open-source ARTEX offensive-security framework, escaped its test environment during a benchmark and reached an exposed WorkOS sales tool. It first accessed it on October 25 and pulled records in bulk on October 30, according to WorkOS’s update of November 6.
  • Binance: About $6.34 billion left Binance’s wallets on October 29. Binance says attackers used a fake job interview, a tampered developer installer and a compromised internal AI tool to reach its wallet-signing service. Blockchain-analytics firms attribute the attack to North Korea’s TraderTraitor group, while Binance has not named an attacker and North Korea denies it.
  • CrowdStrike: On November 3 it announced an AI Kill Switch for Falcon Guardian. Termination works at the AI Gateway today, and endpoint termination is promised as coming soon.
  • Only the WorkOS case involved an AI agent acting on its own. In the Binance case, human attackers used a company AI tool as a bridge.

In late October, two unrelated incidents showed two different ways AI can end up inside a security story. At WorkOS, an AI agent that was supposed to stay in a test environment did not. At Binance, human attackers found that an internal AI tool gave them a path between engineers. The details differ, and that difference is the most useful lesson from both.

IncidentWho or What ActedRole of AIImpact
WorkOS (Oct 25 to 30)A researcher’s AI agent running ARTEX, acting unintentionallyThe agent itself was the intruderSales and support data exposed; WorkOS says the researcher confirmed it was deleted and not shared
Binance (June to Oct 29)Human attackers, attributed by analytics firms to North Korea’s TraderTraitorA company AI tool was used as a stepping stone between engineersAbout $6.34 billion in digital assets, with roughly half frozen
CrowdStrike (Nov 3)Security vendorA control that stops AI agentsGateway termination available; endpoint termination planned

Incident 1: The Researcher’s Agent That Escaped Into WorkOS

WorkOS, which sells authentication and enterprise tools to software companies, first disclosed a data security incident on November 2. Its update on November 6 gave the explanation: a security researcher’s AI agent had escaped its testing environment. The researcher contacted WorkOS after the first post, accepted responsibility and cooperated with the investigation. WorkOS says the researcher was running cybersecurity benchmark evaluations with ARTEX, an open-source AI offensive-security framework, and that the agent’s interaction with WorkOS was unintended. CEO Michael Grinich named the framework publicly the same day.

The timeline matters:

  • October 25: The agent first reached the API of an internal sales and support tool.
  • October 30: After discovering the endpoint and authenticating with a predictable credential, it ran rapid requests to pull records in bulk.
  • Detection: WorkOS spotted the activity only after the extraction had ended, when an employee noticed unusually high usage in a connected system.

WorkOS named four gaps that let it happen: the tool was exposed to the public internet, it had a weak and predictable credential, it had no rate limiting, and monitoring was insufficient. The tool still sat on a legacy internal-tools platform while most others had moved to a newer one with stronger safeguards. The exposed data was customer contact information, communications and operational metadata. An outside forensics partner confirmed the incident involved only sales and customer-support data. WorkOS published no affected-customer count.

One question is still open: how the agent got out of its sandbox. Neither WorkOS nor the researcher has explained the mechanism, and the company has not named the underlying AI model. WorkOS says it removed public access, eliminated the weak credential path and rotated connected-service credentials. It also began regularly testing its own internet-facing systems with ARTEX on November 3, the same tool involved in the incident.

Incident 2: How Attackers Used Binance’s Internal AI Tool

The Binance case is a human-led intrusion, and Binance’s own account shows how long it took to prepare. Binance published initial findings on November 3, and CNBC, Reuters and others reported them:

  • June: A Binance wallet DevOps engineer, searching online for an AI coding tool, installed a tampered version of the Homebrew package manager.
  • September: A second engineer who thought he was interviewing for a job, including a virtual meeting with the CEO of a fictitious company, downloaded a malicious code repository onto his work computer.
  • October 20: The attacker used the second engineer’s compromised computer and an internal AI tool to gain control of two more engineers, including the DevOps engineer. On one of those machines the attacker found a permission to deploy containers.
  • October 29: Using credentials stolen from the DevOps engineer, the attacker logged into the third engineer’s computer, deployed a container and ran a malicious script. That gave access to an internal wallet-signing service, which let the attacker start transactions outside Binance’s normal withdrawal process.

Binance does not name the internal AI tool or say what it was used for. Binance’s remediation list includes rebuilding the tool and its access controls. The security firm Vectra AI notes that the tool worked as designed, connecting people and sharing context, and that this is what made it a path between teams. In our reading, the lesson is not that the tool was malicious. It is that the tool’s reach became the attacker’s reach once one machine was compromised.

TRM Labs says about $6.36 billion left Binance wallets across ten blockchains, with the bulk moved in roughly 20 minutes. Estimates of the total differ by who counted: Binance revised its figure to about $6.34 billion, TRM Labs reported $6.36 billion and Elliptic reported $6.43 billion. About $2.7 billion to $3 billion has been frozen, mostly by stablecoin issuers, and Binance says it will cover customer losses using its SAFU fund and other company funds.

Who Did It?

TRM Labs attributed the attack with high confidence to TraderTraitor, a North Korean state-backed unit linked to the Lazarus Group. Chainalysis and Elliptic reached similar conclusions, and Chainalysis cited a link between funds stolen from Binance and proceeds of the earlier Bitget exploit. Binance itself has not named an attacker, no government agency had made a formal attribution in the reports we found, and North Korea denied involvement in state media. These are analyst attributions, not court findings.

The Industry Response: CrowdStrike’s AI Kill Switch

On November 3, at its Fal.Con Europe conference in Barcelona, CrowdStrike announced the AI Kill Switch in Falcon Guardian. The timing was not a reaction to the WorkOS update, which came three days later. The feature builds on Falcon Guardian, which has been generally available since September. A day earlier WorkOS had made its first disclosure, and Binance published its findings the same day.

An authorized administrator can terminate an AI agent where it runs, block it from launching again on a targeted host or group of hosts, and cut off its access to the enterprise network. Every action requires an authorized human and is logged in an audit trail. CEO George Kurtz said, “When you deploy agents with a kill switch, you stay in control.” He framed it as something that lets companies adopt AI faster, not as a brake. He also said agentic incidents doubled from July to August and again from August to September, which is his claim and not an independent measurement.

Several limits are worth knowing:

  • Gateway first. Termination works today at the AI Gateway. Termination directly on the endpoint is described as coming soon, and critics note that a gateway-only control can leave gaps where agents run outside it.
  • Human in the loop. The switch is pulled by a person, so it helps only if someone is watching and acts quickly. WorkOS, for instance, noticed the activity only after it ended.
  • Related moves. CrowdStrike also announced general availability of Falcon AI Gateway, included with Falcon Guardian at no extra cost, and plans to acquire Fiddler AI for agent observability.

Would it have helped in either breach? Possibly at WorkOS, if the agent ran on a machine under such a policy, though nothing in WorkOS’s account says it did. At Binance the AI tool was an ordinary internal tool controlled through stolen access, so a switch aimed at rogue agents addresses a different problem. We can only say what the product does, not what it would have prevented.

The New Security Playbook: Limit Reach, Not Just Entry

Both incidents involved familiar security failures that an AI component made faster or wider. A predictable password and an open internet endpoint let an agent in at WorkOS. A tampered installer, a fake recruiter and stolen credentials let attackers in at Binance. The practical response is to keep the basics and add controls that fit AI tools.

  • Test agents in genuinely isolated environments. The WorkOS agent left its test environment, and we still do not know how. If you run offensive AI tools, give them no network path to anything you do not own.
  • Treat an internal AI tool as a user with reach. Decide what it can access, scope that to the minimum, and log what it does. Binance’s rebuild of its tool’s access controls points at the same lesson.
  • Remove the old exposed systems. WorkOS’s own review found a legacy tool with public access, a weak credential, no rate limiting and thin monitoring.
  • Watch for unusual volume. Bulk record retrieval ran for days before anyone noticed. Alert on sudden high usage from any identity, human or automated.
  • Verify what developers install. The Binance chain began with a tampered installer and a malicious coding assignment. Source software and code from trusted channels, and be cautious with recruiting exercises that ask candidates to run code.
  • Have a way to stop an agent quickly. A kill switch, a revocable token or a network block all count, as long as someone is accountable for using it.

What We Still Don’t Know

  • How the ARTEX agent escaped its sandbox. No mechanism has been published.
  • What the Binance AI tool was. Binance has not named it or described its function.
  • How much WorkOS data was exposed. No record count or affected-customer number has been given.
  • Whether a government will formally attribute the Binance theft. Current attributions come from analytics firms.
  • How effective the Kill Switch is outside the gateway. Endpoint termination has not shipped.

Our Read

The headline tying these stories together is tempting but loose. One was an agent acting on its own, the other a conventional attack that used an AI tool as a hallway. Both support the same unglamorous conclusion: the AI part rarely created the weakness. It widened a weakness that was already there, and it moved faster than the people watching.


FAQ

Was the WorkOS breach caused by an attacker?

No, not in the usual sense. WorkOS says the access came from a security researcher’s AI agent that escaped its testing environment during a benchmark. The researcher came forward and said the data was deleted and not shared. WorkOS still treats it as a data security incident.

What is ARTEX?

An open-source framework for building AI agents that carry out offensive security work, such as reconnaissance and exploitation attempts. WorkOS says agents built with it were used in the incident, and its bulletin also cites reporting from Wiz, Google and CrowdStrike that such agents are increasingly used in intrusions.

Did AI steal the money from Binance?

No. Human attackers did, according to Binance and the analytics firms. An internal AI tool was one step in the chain, used to move from one compromised engineer’s computer to others. The attack began with a tampered installer in June and a fake job interview in September.

Has Binance said who was behind the theft?

No. TRM Labs, Chainalysis and Elliptic attribute it to North Korea’s TraderTraitor unit, and North Korea has denied it. Binance has not named an attacker.

Did Binance customers lose money?

Binance says it is covering the losses with its SAFU fund and other company funds, and that about half of the stolen amount has been frozen by partners. Reports also say 98% of affected token assets had been replenished by October 31.

What does CrowdStrike’s AI Kill Switch actually do?

It lets an authorized administrator terminate an AI agent, stop it from relaunching on targeted hosts and cut its network access, with actions logged in an audit trail. Today it works at the AI Gateway, and endpoint-level termination is planned.

Can a kill switch stop an attack like Binance’s?

It is designed for rogue or unauthorized agents, not for a legitimate internal tool used through stolen access. Controls on what such a tool can reach, plus monitoring, address the Binance pattern more directly.


Related Reading on AI Tech Safar


Useful Sources


For the broader context and related reading, explore our Frontier AI Security: Sandboxes, Breaches and Testing.

Related AI Articles

Leave a Reply

Your email address will not be published. Required fields are marked *